| The rapid development of hardware in recent years has continuously improved its computing capabilities.The new neural network algorithm in deep learning has led to the rapid development of target detection,and the real-time performance and accuracy have been greatly improved.However,for vehicle and pedestrian multi-target detection scenarios,the first model relies on powerful computing power;the second model has a problem of flickering between adjacent frames in the video detection;the third model will have a lot of problems on the test side.Redundant data.This will bring great troubles to the practical application of vehicle and pedestrian detection.Therefore,optimizing the detection algorithm for vehicles and pedestrians has both theoretical and application value for solving the above problems.To solve the above problems,a multi-target detection method based on lightweight neural network(Ghost-DW-BIFPN)is proposed.First,the data set was expanded through the corresponding data set expansion technology,and the more stable K-Means++ algorithm was re-selected for clustering to replace the original K-Means algorithm;secondly,for the problem of larger models and more parameters,the network architecture Reconstruction,with the goal of lightweight,using Ghost Net to improve the backbone network,the feature fusion module is improved on the basis of BIFPN with excellent fusion effect,and a new bidirectional feature pyramid network based on deep convolution(DW-BIFPN)is generated,combining the two parts The improvement is combined into a new target detection algorithm;then for the drawbacks of frame-by-frame processing,frame skipping is used to accelerate the real-time video processing,and for the problem of candidate frame flicker in video detection,the Deep SORT tracking algorithm is used to optimize processing,combined with frame skipping method to effectively optimize the problems in video detection.Finally,for the urban road traffic scene,an optimization algorithm for extracting the region of interest using lane line information detection is proposed,which realizes the detection method of vehicles and pedestrians only in the range of the region of interest.Experiments show that the improved backbone network structure has achieved better feature extraction results,and the final network structure is reduced to about 10% in network parameters while still obtaining high accuracy;the use of Deep SORT algorithm to optimize the candidate frame flicker phenomenon;finally The region of interest obtained by using the lane information effectively reduces time consumption by approximately 33%. |
